Revolut Reports Large Coding-Inference Savings With Self-Hosted GLM

Revolut says self-hosting GLM on 56 NVIDIA B200s reduced coding inference costs to about one-tenth and improved engineering metrics. Coverage suggests abundant, inexpensive tokens—not simply a smarter model—may explain the gains, while Claude Code and Codex remain in use for harder tasks. The results are company-reported and lack independent methodology, so their productivity impact and scalability remain to be tested.